• that many residuals can be computed for GLM and that they are often useless
• that computing residuals by parametric bootstraps in the way out
• that traditional goodness of fit tests are bad
• that most assumptions behing GLM are similar to those for LM
• that General Additive Models (GAM) can be useful to improve linearity
• how to diagnose and tackle overdispersion, zero-augmentation and separation

set.seed(1L)
Aliens <- simulate_Aliens_GLM()
mod_gauss <- glm(size  ~ humans_eaten, family = gaussian(), data = Aliens)
mod_poiss <- glm(eggs  ~ humans_eaten, family = poisson(),  data = Aliens)
mod_binar <- glm(happy ~ humans_eaten, family = binomial(), data = Aliens)
mod_binom <- glm(cbind(blue_eyes, pink_eyes) ~ humans_eaten, family = binomial(), data = Aliens)

par(mfrow = c(2, 2))
plot(mod_poiss)

par(mfrow = c(2, 2))
plot(mod_binar)

par(mfrow = c(2, 2))
plot(mod_binom)

rbind(
  residuals(mod_poiss, type = "deviance")[1:2],
  residuals(mod_poiss, type = "pearson")[1:2],
  residuals(mod_poiss, type = "working")[1:2],
  residuals(mod_poiss, type = "response")[1:2])

##              1          2
## [1,] 0.2639365 -1.2774761
## [2,] 0.2769535 -0.9033120
## [3,] 0.3179479 -1.0000000
## [4,] 0.2412447 -0.8159726

rbind(
  residuals(mod_binom, type = "deviance")[1:2],
  residuals(mod_binom, type = "pearson")[1:2],
  residuals(mod_binom, type = "working")[1:2],
  residuals(mod_binom, type = "response")[1:2])

##              1          2
## [1,] 1.4087111 -0.8875765
## [2,] 1.3994400 -0.9179635
## [3,] 1.0629762 -0.6653824
## [4,] 0.2632007 -0.1407139

Note 1: we can also compute partial residuals, which we shall see later.

Note 2: this is also true for LM and Gaussian GLM, but it makes no difference in such case (except for partial residuals).

They are not necessarily normaly distributed!

These are the residuals that we saw in LM: observed - predicted. Simple but pretty much useless.

rbind(residuals(mod_gauss, type = "response")[1:2],
      (mod_gauss$y - mod_gauss$fitted.values)[1:2])

##             1         2
## [1,] 2.434285 -2.804164
## [2,] 2.434285 -2.804164

rbind(residuals(mod_poiss, type = "response")[1:2],
      (mod_poiss$y - mod_poiss$fitted.values)[1:2])

##              1          2
## [1,] 0.2412447 -0.8159726
## [2,] 0.2412447 -0.8159726

rbind(residuals(mod_binom, type = "response")[1:2],
      (mod_binom$y - mod_binom$fitted.values)[1:2])

##              1          2
## [1,] 0.2632007 -0.1407139
## [2,] 0.2632007 -0.1407139

These are response residuals scaled by the square root of the variance function and accounting for their prior weights.

variances <- poisson()$variance(mod_poiss$fitted.values)
rbind(residuals(mod_poiss, type = "pearson")[1:2],
      (residuals(mod_poiss, type = "response") * sqrt(mod_poiss$prior.weights / variances))[1:2])

##              1         2
## [1,] 0.2769535 -0.903312
## [2,] 0.2769535 -0.903312

variances <- binomial()$variance(mod_binom$fitted.values)
rbind(residuals(mod_binom, type = "pearson")[1:2],
      (residuals(mod_binom, type = "response") * sqrt(mod_binom$prior.weights / variances))[1:2])

##            1          2
## [1,] 1.39944 -0.9179635
## [2,] 1.39944 -0.9179635

These residuals are often use to check the quality of the fit.

The sum of the squared Pearson residuals gives the Pearson \(\chi^2\) statistic of goodness of fit:

(X <- sum(residuals(mod_binom, type = "pearson")^2))

## [1] 88.77233

This statistic is sometimes used to test the goodness of fit of the model or to test for overdispersion, but since its usage is often not optimal (e.g. it requires a lot of replications within each grouping) and the power if this test is poor, I would not recommend you to use such test.

1 - pchisq(X, mod_poiss$df.residual)  ## Pearson goodness of fit test

## [1] 0.7366322

X / mod_poiss$df.residual  ## measure of overdispersion

## [1] 0.9058401

These are the residuals that are minimized after the fit.

Contrary to those computed by family()$dev.resids they are here not squared.

residuals(mod_poiss, type = "deviance")[1:2]

##          1          2 
##  0.2639365 -1.2774761

rbind((residuals(mod_poiss, type = "deviance")[1:2])^2,
      with(mod_poiss, poisson()$dev.resids(y = y, mu = fitted.values, wt = prior.weights))[1:2])

##               1        2
## [1,] 0.06966248 1.631945
## [2,] 0.06966248 1.631945

The sum of the squared deviance residuals gives the residual deviance of the model fit:

c(sum(residuals(mod_gauss, type = "deviance")^2), deviance(mod_gauss))

## [1] 2172.918 2172.918

c(sum(residuals(mod_poiss, type = "deviance")^2), deviance(mod_poiss))

## [1] 116.8588 116.8588

c(sum(residuals(mod_binom, type = "deviance")^2), deviance(mod_binom))

## [1] 92.83338 92.83338

These residuals are sometimes also used to check the quality of the fit, in the exact same fashion as the Pearson's residuals, but again (and for the same reasons), I would not recommend you to do that.

For mod_poiss

p <- replicate(1000, {
  d <- simulate_Aliens_GLM()
  m <- update(mod_poiss, data = d)
  X <- sum(residuals(m, type = "pearson")^2)
  1 - pchisq(X, m$df.residual)
})
plot(ecdf(p))
abline(0, 1, lty = 2, col = "green")

p <- replicate(1000, {
  d <- simulate_Aliens_GLM()
  m <- update(mod_poiss, data = d)
  dev <- deviance(m)
  1 - pchisq(dev, m$df.residual)
})
plot(ecdf(p))
abline(0, 1, lty = 2, col = "green")

For mod_binar

p <- replicate(1000, {
  d <- simulate_Aliens_GLM()
  m <- update(mod_binar, data = d)
  X <- sum(residuals(m, type = "pearson")^2)
  1 - pchisq(X, m$df.residual)
})
plot(ecdf(p))
abline(0, 1, lty = 2, col = "green")

p <- replicate(1000, {
  d <- simulate_Aliens_GLM()
  m <- update(mod_binar, data = d)
  dev <- deviance(m)
  1 - pchisq(dev, m$df.residual)
})
plot(ecdf(p))
abline(0, 1, lty = 2, col = "green")

For mod_binom

p <- replicate(1000, {
  d <- simulate_Aliens_GLM()
  m <- update(mod_binom, data = d)
  X <- sum(residuals(m, type = "pearson")^2)
  1 - pchisq(X, m$df.residual)
})
plot(ecdf(p))
abline(0, 1, lty = 2, col = "green")

p <- replicate(1000, {
  d <- simulate_Aliens_GLM()
  m <- update(mod_binom, data = d)
  dev <- deviance(m)
  1 - pchisq(dev, m$df.residual)
})
plot(ecdf(p))
abline(0, 1, lty = 2, col = "green")

These are the residuals used during the last iteration of the iterative fitting procedure. Useless otherwise.

rbind(mod_poiss$residuals[1:2],
      ((mod_poiss$y - mod_poiss$fitted.values)/poisson()$mu.eta(mod_poiss$linear.predictors))[1:2])

##              1  2
## [1,] 0.3179479 -1
## [2,] 0.3179479 -1

rbind(mod_binom$residuals[1:2],
      ((mod_binom$y - mod_binom$fitted.values)/binomial()$mu.eta(mod_binom$linear.predictors))[1:2])

##             1          2
## [1,] 1.062976 -0.6653824
## [2,] 1.062976 -0.6653824

These are residuals expressed at the level of each predictor.

mod_UK_small <- glm(milk ~ drink + sex + cigarettes, data = UK[1:10, ], family = poisson())
residuals(mod_UK_small, type = "partial")

##         drink        sex cigarettes
## 1    1.164787 -1.2462280 -1.3057428
## 2    1.452278 -0.1985045 -0.4043752
## 3    2.369917  0.5129145  0.3070438
## 4    1.164787 -1.2462280 -1.3057428
## 6    2.598608  0.9478250  0.6400400
## 7    1.890839 -0.3139566  0.6456712
## 8    4.317480  1.9064653  1.8469505
## 9  -17.699639 -1.2462280 -0.6942572
## 10   1.768674 -0.0883280 -0.7018558
## attr(,"constant")
## [1] -1.662461

These are residuals expressed at the level of each predictor.

Computation:

(p <- predict(mod_UK_small, type = "terms")) ## prediction expressed per predictor

##         drink       sex cigarettes
## 1    2.164787 -0.246228 -0.3057428
## 2    1.958568  0.307785  0.1019143
## 3    2.164787  0.307785  0.1019143
## 4    2.164787 -0.246228 -0.3057428
## 6    1.958568  0.307785  0.0000000
## 7    1.958568 -0.246228  0.7133998
## 8    2.164787 -0.246228 -0.3057428
## 9  -16.699639 -0.246228  0.3057428
## 10   2.164787  0.307785 -0.3057428
## attr(,"constant")
## [1] -1.662461

c(sum(p[1, ]) + attr(p, "constant"), predict(mod_UK_small, type = "link")[1])

##                       1 
## -0.04964481 -0.04964481

These are residuals expressed at the level of each predictor.

Computation:

rbind((p + residuals(mod_UK_small, type = "working"))[1, ],
      residuals(mod_UK_small, type = "partial")[1, ])

##         drink       sex cigarettes
## [1,] 1.164787 -1.246228  -1.305743
## [2,] 1.164787 -1.246228  -1.305743

Partial residuals could be useful to check for departure from linearity (if they were not computed on the basis of working residuals…).

It is probably more useful for LM in which working residuals and response residuals are the same.

library(car)
par(mfrow = c(1, 3))
crPlots(mod_UK_small, terms = ~ drink)
crPlots(mod_UK_small, terms = ~ sex)
crPlots(mod_UK_small, terms = ~ cigarettes)

mod_UK <- glm(milk ~ drink + sex + cigarettes, data = UK, family = poisson())
par(mfrow = c(1, 3))
crPlots(mod_UK, terms = ~ drink)
crPlots(mod_UK, terms = ~ sex)
crPlots(mod_UK, terms = ~ cigarettes)

set.seed(1L)
s <- simulate(mod_poiss, 1000)
r <- sapply(s, function(i) i + runif(nrow(mod_poiss$model), min = -0.5, max = 0.5))
hist(r[1, ], main = "distrib of first fitted value", nclass = 30)
abline(v = mod_poiss$y[1] + runif(1, min = -0.5, max = 0.5), col = "red", lwd = 2, lty = 2)

plot(ecdf1 <- ecdf(r[1, ]), main = "cdf of first fitted value")
noise <- runif(1, min = -0.5, max = 0.5)
simulated_residual_1 <-  ecdf1(mod_poiss$y[1] + noise)
segments(x0 = mod_poiss$y[1] + noise, y0 = 0, y1 =  simulated_residual_1, col = "red", lwd = 2)
arrows(x0 = mod_poiss$y[1] + noise, x1 = -1, y0 = simulated_residual_1, col = "red", lwd = 2)

plot(ecdf2 <- ecdf(r[2, ]), main = "cdf of second fitted value")
noise <- runif(1, min = -0.5, max = 0.5)
simulated_residual_2 <-  ecdf2(mod_poiss$y[2] + noise)
segments(x0 = mod_poiss$y[2] + noise, y0 = 0, y1 =  simulated_residual_2, col = "red", lwd = 2)
arrows(x0 = mod_poiss$y[2] + noise, x1 = -1, y0 = simulated_residual_2, col = "red", lwd = 2)

simulated_residuals <- rep(NA, nrow(mod_poiss$model))
for (i in 1:nrow(mod_poiss$model)) {
  ecdf_fn <- ecdf(r[i, ])
  simulated_residuals[i] <- ecdf_fn(mod_poiss$y[i] + runif(1, min = -0.5, max = 0.5))
}

plot(simulated_residuals)

plot(ecdf(simulated_residuals))

library(DHARMa)
mod_poiss_simres <- simulateResiduals(mod_poiss)
plot(mod_poiss_simres)

mod_binar_simres <- simulateResiduals(mod_binar)
plot(mod_binar_simres)

mod_binom_simres <- simulateResiduals(mod_binom)
plot(mod_binom_simres)

mod_UK_simres <- simulateResiduals(mod_UK)
plot(mod_UK_simres)

Model structure

• linearity (for the linear predictor)
• lack of perfect multicollinearity (design matrix of full rank)
• predictor variables have fixed values

Errors

• independence (no serial autocorrelation)
• lack of overdispersion and underdispersion

It is very difficult…

but we may try to

• plot partial residuals
• plot simulated residuals & run an uniformity test
• plot the predictions from a GAM model

set.seed(1L)
Aliens2 <- data.frame(humans_eaten = runif(100, min = 0, max = 15))
Aliens2$eggs <- rpois( n = 100, lambda = exp(1 + 0.2 * Aliens2$humans_eaten - 0.02 *  Aliens2$humans_eaten^2))
mod_poiss2bad <- glm(eggs ~ humans_eaten, data = Aliens2, family = poisson()) ## mispecified model
(mod_poiss2good <- glm(eggs ~ poly(humans_eaten, 2, raw = TRUE), data = Aliens2, family = poisson())) ## good model

## 
## Call:  glm(formula = eggs ~ poly(humans_eaten, 2, raw = TRUE), family = poisson(), 
##     data = Aliens2)
## 
## Coefficients:
##                        (Intercept)  poly(humans_eaten, 2, raw = TRUE)1  poly(humans_eaten, 2, raw = TRUE)2  
##                            1.18654                             0.13015                            -0.01463  
## 
## Degrees of Freedom: 99 Total (i.e. Null);  97 Residual
## Null Deviance:       137.5 
## Residual Deviance: 95.12     AIC: 376.7

plot(I(1 + 0.2 * Aliens2$humans_eaten - 0.02 *  Aliens2$humans_eaten^2) ~ Aliens2$humans_eaten,
     ylab = "eta", ylim = c(-1, 2))
data.for.pred <- data.frame(humans_eaten = 0:15)
points(predict(mod_poiss2good, newdata = data.for.pred) ~ I(0:15), col = "blue")
points(predict(mod_poiss2bad, newdata = data.for.pred) ~ I(0:15), col = "red")

plot(exp(1 + 0.2 * Aliens2$humans_eaten - 0.02 *  Aliens2$humans_eaten^2) ~ Aliens2$humans_eaten,
     ylab = "predicted number of eggs", ylim = c(0, 6))
points(predict(mod_poiss2good, newdata = data.for.pred, type = "response") ~ I(0:15), col = "blue")
points(predict(mod_poiss2bad, newdata = data.for.pred, type = "response") ~ I(0:15), col = "red")

crPlots(mod_poiss2bad, terms = ~ humans_eaten)

crPlots(mod_poiss, terms = ~ humans_eaten)  ## cannot do that for mod_poiss2good

plot(s_bad <- simulateResiduals(mod_poiss2bad))

plot(s_good <- simulateResiduals(mod_poiss2good))

The lack of linearity is unfortunatly not detected here:

testUniformity(s_bad)

## 
##  One-sample Kolmogorov-Smirnov test
## 
## data:  simulationOutput$scaledResiduals
## D = 0.072, p-value = 0.6777
## alternative hypothesis: two-sided

testUniformity(s_good)

## 
##  One-sample Kolmogorov-Smirnov test
## 
## data:  simulationOutput$scaledResiduals
## D = 0.062, p-value = 0.8367
## alternative hypothesis: two-sided

library(mgcv)
mod_poiss2GAM <- gam(eggs ~ s(humans_eaten), data = Aliens2, family = poisson())
plot(mod_poiss2GAM)

You can run the Durbin Watson test on the simulated residuals:

testTemporalAutocorrelation(s_good, time = mod_poiss2good$fitted.values, plot = FALSE)

## 
##  Durbin-Watson test
## 
## data:  simulationOutput$scaledResiduals ~ 1
## DW = 2.2265, p-value = 0.8737
## alternative hypothesis: true autocorrelation is greater than 0

You can run the Durbin Watson test on the simulated residuals:

testTemporalAutocorrelation(s_good, time = Aliens2$humans_eaten, plot = FALSE)

## 
##  Durbin-Watson test
## 
## data:  simulationOutput$scaledResiduals ~ 1
## DW = 2.2876, p-value = 0.9268
## alternative hypothesis: true autocorrelation is greater than 0

Note: As for LM it is good practice to test for the lack of serial autocorrelation along all your predictors and not just along the fitted values.

Survival times series that are discrete and complete can be analysed using a binomial (binary!) GLM.

Example: the study of the influence of rainfall on mortality (here A dies at 5 yrs old and B at 3 yrs old).

data.frame(age = c(1:5, 1:5),
      id = c(rep("A", 5), rep("B", 5)),
      death = c(0, 0, 0, 0, 1, 0, 0, 1, NA, NA),
      annual_rain = c(100, 120, 310, 50, 200, 45, 100, 320, 100, 120))

##    age id death annual_rain
## 1    1  A     0         100
## 2    2  A     0         120
## 3    3  A     0         310
## 4    4  A     0          50
## 5    5  A     1         200
## 6    1  B     0          45
## 7    2  B     0         100
## 8    3  B     1         320
## 9    4  B    NA         100
## 10   5  B    NA         120

In such a case, a random effect for the identity of the individual is not needed! So there is no need for mixed model if individuals are all equally independent from each other.

A GLM assumes a particular relationship between mu and var(mu)

p <- seq(0, 1, 0.1)
lambda <- 0:10
theta <- 0:10
v_b <- binomial()$variance(p)
v_p <- poisson()$variance(lambda)
v_G <- Gamma()$variance(theta)
par(mfrow = c(1, 3), las = 2)
plot(v_b ~ p); plot(v_p ~ lambda); plot(v_G ~ theta)

Overdispersion = more variance than expected

• very common \(\rightarrow\) increases false positive
• specially relevant for Poisson and Binomial
• irrelevant for the binary case! (don't look for it)

Usual suspects:

• lack of linearity
• unobserved heterogeneity
• zero-augmentation

Underdispersion = less variance than expected

• rather rare \(\rightarrow\) increases false negative

A toy example

set.seed(1L)
popA <- data.frame(humans_eaten = runif(50, min = 0, max = 15))
popA$eggs <- rpois(n = 50, lambda = exp(-1 + 0.05 * popA$humans_eaten))
popA$pop <- "A"
popB <- data.frame(humans_eaten = runif(50, min = 0, max = 15))
popB$eggs <- rpois(n = 50, lambda = exp(-3 + 0.4 * popB$humans_eaten))
popB$pop <- "B"
AliensMix <- rbind(popA, popB)
(mod_poissMix <- glm(eggs ~ humans_eaten, family = poisson(), data = AliensMix))

## 
## Call:  glm(formula = eggs ~ humans_eaten, family = poisson(), data = AliensMix)
## 
## Coefficients:
##  (Intercept)  humans_eaten  
##      -3.3414        0.3756  
## 
## Degrees of Freedom: 99 Total (i.e. Null);  98 Residual
## Null Deviance:       465.7 
## Residual Deviance: 214.2     AIC: 344.9

How to test it?

A widely used not so good way:

cbind(disp = mod_poissMix$deviance / mod_poissMix$df.residual,
      pv = 1 - pchisq(mod_poissMix$deviance, mod_poissMix$df.residual))

##          disp           pv
## [1,] 2.185731 1.189718e-10

A slightly better way:

cbind(disp = sum(residuals(mod_poissMix, type = "pearson")^2) / mod_poissMix$df.residual,
      pv = 1 - pchisq(sum(residuals(mod_poissMix, type = "pearson")^2), mod_poissMix$df.residual))

##          disp           pv
## [1,] 2.267855 1.215561e-11

How to test it?

A better way (?)

r <- simulateResiduals(mod_poissMix, refit = TRUE)
testOverdispersion(r)

## 
##  DHARMa nonparametric overdispersion test via comparison to simulation under H0 = fitted model
## 
## data:  r
## dispersion = 2.3112, p-value < 2.2e-16
## alternative hypothesis: overdispersion

• fix linearity issues
• fix heterogeneity issues (if you have the data)

mod_poissMix2 <- glm(eggs ~ pop*humans_eaten, family = poisson(), data = AliensMix)
r2 <- simulateResiduals(mod_poissMix2, refit = TRUE)
testOverdispersion(r2)  ## you can change options to test for underdispersion

## 
##  DHARMa nonparametric overdispersion test via comparison to simulation under H0 = fitted model
## 
## data:  r2
## dispersion = 0.7602, p-value = 0.944
## alternative hypothesis: overdispersion

• fix linearity issues
• fix heterogeneity issues (if you have the data)
• model the overdispersion
• try another probability distribution
• there are specific solutions if the origin is zero-augmentation

For Poisson

Variance = \(k \times \mu\)

mod_poissMixQ <- glm(eggs ~ humans_eaten, family = quasipoisson(), data = AliensMix)
summary(mod_poissMixQ)$coef

##               Estimate Std. Error   t value     Pr(>|t|)
## (Intercept)  -3.341449 0.54852447 -6.091705 2.195748e-08
## humans_eaten  0.375603 0.04364544  8.605779 1.272120e-13

Anova(mod_poissMixQ, test = "F")

## Analysis of Deviance Table (Type II tests)
## 
## Response: eggs
## Error estimate based on Pearson residuals 
## 
##              Sum Sq Df F value    Pr(>F)    
## humans_eaten 251.54  1  110.92 < 2.2e-16 ***
## Residuals    222.25 98                      
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

For Poisson

Variance = \(k \times \mu\)

It works, but we cannot do much with these type of fit:

logLik(mod_poissMixQ)

## 'log Lik.' NA (df=2)

For binomial

• irrelevant for the binary case!
• there is also quasibinomial(), with the same limit (no likelihood)
• can happen in the general case

Solution

• add a random effect with one level per individual (see GLMM)

• Poisson (\(V(\mu) = \mu\))
• Negative binomial (\(V(\mu) = \mu + \frac{\mu^2}{\theta}\), if \(\theta = 1\) this is the geometric model)
• Conway-Maxwell-Poisson (\(V(\mu) =\) something complex)

For comparison

mod_poissMix <- glm(eggs ~ humans_eaten, family = poisson(), data = AliensMix)
sum(residuals(mod_poissMix, type = "pearson")^2) / mod_poissMix$df.residual

## [1] 2.267855

summary(mod_poissMix)$coefficients

##               Estimate Std. Error   z value     Pr(>|z|)
## (Intercept)  -3.341449 0.36421196 -9.174463 4.538303e-20
## humans_eaten  0.375603 0.02897991 12.960806 2.040928e-38

library(MASS)
mod_poissMixNB <- glm.nb(eggs ~ humans_eaten, data = AliensMix)
sum(residuals(mod_poissMixNB, type = "pearson")^2) / mod_poissMix$df.residual

## [1] 0.836146

summary(mod_poissMixNB)$coefficients

##                Estimate Std. Error   z value     Pr(>|z|)
## (Intercept)  -2.8414180 0.47988075 -5.921092 3.198110e-09
## humans_eaten  0.3292865 0.04508882  7.303063 2.812901e-13

c(AIC(mod_poissMix), AIC(mod_poissMixNB), logLik(mod_poissMixNB))

## [1]  344.8726  280.7114 -137.3557

negbin <- spaMM::negbin ## otherwise spaMM tries to use negbin from mgcv, which won't work
mod_poissMixSpaMM   <- fitme(eggs ~ humans_eaten, family = poisson(), data = AliensMix)
mod_poissMixNBSpaMM <- fitme(eggs ~ humans_eaten, family = negbin(), data = AliensMix)
summary(mod_poissMixNBSpaMM)

## formula: eggs ~ humans_eaten
## ML: Estimation of corrPars and NB_shape by ML.
##     Estimation of fixed effects by ML.
## Estimation of NB_shape by 'outer' ML, maximizing p_v.
## Family: Neg.binomial(shape=1.017) ( link = log ) 
##  ------------ Fixed effects (beta) ------------
##              Estimate Cond. SE t-value
## (Intercept)   -2.8414  0.47988  -5.921
## humans_eaten   0.3293  0.04509   7.303
##  ------------- Likelihood values  -------------
##                         logLik
## p(h)   (Likelihood): -137.3557

c(AIC(mod_poissMixSpaMM), AIC(mod_poissMixNBSpaMM), logLik(mod_poissMixNBSpaMM))

##        marginal AIC:        marginal AIC:                  p_v 
##             344.8726             280.7114            -137.3557

It can fit both over- and under- dispersion!

• overdispersion: nu < 1 (nu = 0 corresponds to the geometric distribution)
• expected dispersion (i.e. Poisson): nu = 1
• underdispersion: nu > 1

Replicating Poisson:

mod_poissMixCP <- glm(eggs ~ humans_eaten, family = COMPoisson(nu = 1), data = AliensMix)
mod_poissMixCP$coef

##  (Intercept) humans_eaten 
##   -3.8361936    0.4030429

c(AIC(mod_poissMix), AIC(mod_poissMixNB), AIC(mod_poissMixCP))

## [1] 344.8726 280.7114 382.7765

It can fit both over- and under- dispersion!

mod_poissMixCPSpaMM <- fitme(eggs ~ humans_eaten, family = COMPoisson(), data = AliensMix)
summary(mod_poissMixCPSpaMM)

## formula: eggs ~ humans_eaten
## ML: Estimation of corrPars and COMP_nu by ML.
##     Estimation of fixed effects by ML.
## Estimation of COMP_nu by 'outer' ML, maximizing p_v.
## Family: COMPoisson(nu=0.05007) ( link = loglambda ) 
##  ------------ Fixed effects (beta) ------------
##              Estimate Cond. SE t-value
## (Intercept)   -2.1584  0.24533  -8.798
## humans_eaten   0.1502  0.01743   8.613
##  ------------- Likelihood values  -------------
##                         logLik
## p(h)   (Likelihood): -157.0442

c(AIC(mod_poissMixSpaMM), AIC(mod_poissMixNBSpaMM), AIC(mod_poissMixCP), AIC(mod_poissMixCPSpaMM))

##        marginal AIC:        marginal AIC:                             marginal AIC: 
##             344.8726             280.7114             382.7765             318.0884

Note: here we estimate nu, so it can take (much) longer time to fit!

d <- data.frame(humans_eaten = seq(0, 15, length = 1000))
p <- predict(mod_poissMixSpaMM, newdata = d, intervals = "predVar")
plot(p ~ d$humans_eaten, ylim = range(mod_poissMixSpaMM$data$eggs), type = "l", lwd = 3)
points(attr(p, "interval")[, 1] ~ d$humans_eaten, lty = 2, type = "l")
points(attr(p, "interval")[, 2] ~ d$humans_eaten, lty = 2, type = "l")
p <- predict(mod_poissMixNBSpaMM, newdata = d, intervals = "predVar")
points(p ~ d$humans_eaten, type = "l", col = "red", lwd = 3)
points(attr(p, "interval")[, 1] ~ d$humans_eaten, lty = 2, type = "l", col = "red")
points(attr(p, "interval")[, 2] ~ d$humans_eaten, lty = 2, type = "l", col = "red")
p <- predict(mod_poissMixCPSpaMM, newdata = d, intervals = "predVar")
points(p ~ d$humans_eaten, type = "l", col = "blue", lwd = 3)
points(attr(p, "interval")[, 1] ~ d$humans_eaten, lty = 2, type = "l", col = "blue")
points(attr(p, "interval")[, 2] ~ d$humans_eaten, lty = 2, type = "l", col = "blue")
legend("topleft", fill = c("black", "red", "blue"),
       legend = c("Poisson", "Negative Binomial", "Conway-Maxwell-Poisson "), bty = "n")

What is it?

It occurs for binomial or Poisson events when too many zeros occur.

Why does it occur?

It can occur when the response results from a 2 (or more) steps process

Examples:

• detection issue (low counts are less detected, e.g. counting cells on microscope)
• biological (e.g. infection, then spread of microbes)

You may try again to use the negative binomial or the COM-Poisson distribution, but an alternative that is often best is to combine two models (during the fitting procedure, not sequentially)!

We have 2 main options for this:

Fit an hurdle model

• binomial (or truncated count distribution) + truncated Poisson or truncated negative binomial
• a single source of zeros
• e.g. number of offspring

Fit a zero-inflation model

• binomial (or truncated count distribution) + Poisson or negative binomial
• two sources of zeros
• e.g. number of viruses in individuals (0 for unexposed, 0 for exposed with strong immune system)

If you are not sure where the zeros come from, try both!

set.seed(1L)
AliensZ <- simulate_Aliens_GLM(N = 1000)
AliensZ$eggs <- AliensZ$eggs * AliensZ$happy ## unhappy Aliens loose their eggs :-(
barplot(table(AliensZ$eggs))

mod_Zpoiss <- glm(eggs ~ humans_eaten, data = AliensZ, family = poisson())
r <- simulateResiduals(mod_Zpoiss)
testZeroInflation(r, plot = FALSE)

## 
##  DHARMa zero-inflation test via comparison to expected zeros with simulation under H0 = fitted model
## 
## data:  r
## ratioObsExp = 1.0723, p-value < 2.2e-16
## alternative hypothesis: more

mean(sum(AliensZ$eggs == 0) / sum(dpois(0, fitted(mod_Zpoiss))))

## [1] 1.072326

testZeroInflation(r, plot = TRUE)

## 
##  DHARMa zero-inflation test via comparison to expected zeros with simulation under H0 = fitted model
## 
## data:  r
## ratioObsExp = 1.0723, p-value < 2.2e-16
## alternative hypothesis: more

mod_Znb <- glm.nb(eggs ~ humans_eaten, data = AliensZ)
r <- simulateResiduals(mod_Znb)
testZeroInflation(r, plot = FALSE)

## 
##  DHARMa zero-inflation test via comparison to expected zeros with simulation under H0 = fitted model
## 
## data:  r
## ratioObsExp = 1.0112, p-value = 0.252
## alternative hypothesis: more

testZeroInflation(r, plot = TRUE)

## 
##  DHARMa zero-inflation test via comparison to expected zeros with simulation under H0 = fitted model
## 
## data:  r
## ratioObsExp = 1.0112, p-value = 0.252
## alternative hypothesis: more

library(pscl)
mod_Zhurd1 <- hurdle(eggs ~ humans_eaten | humans_eaten, dist = "poisson", zero.dist = "binomial",
                     data = AliensZ)
mod_Zhurd2 <- hurdle(eggs ~ humans_eaten | 1, dist = "poisson", zero.dist = "binomial", data = AliensZ)
lmtest::lrtest(mod_Zhurd1, mod_Zhurd2)

## Likelihood ratio test
## 
## Model 1: eggs ~ humans_eaten | humans_eaten
## Model 2: eggs ~ humans_eaten | 1
##   #Df  LogLik Df  Chisq Pr(>Chisq)    
## 1   4 -703.87                         
## 2   3 -818.64 -1 229.53  < 2.2e-16 ***
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

mean(sum(AliensZ$eggs == 0) / sum(predict(mod_Zhurd1, type = "prob")[, 1]))

## [1] 1

mod_Zzi1 <- zeroinfl(eggs ~ humans_eaten | humans_eaten, dist = "poisson", data = AliensZ)
mod_Zzi2 <- zeroinfl(eggs ~ humans_eaten | 1, dist = "poisson", data = AliensZ)
lmtest::lrtest(mod_Zzi1, mod_Zzi2)

## Likelihood ratio test
## 
## Model 1: eggs ~ humans_eaten | humans_eaten
## Model 2: eggs ~ humans_eaten | 1
##   #Df  LogLik Df  Chisq Pr(>Chisq)    
## 1   4 -703.46                         
## 2   3 -717.75 -1 28.586  8.961e-08 ***
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

mean(sum(AliensZ$eggs == 0) / sum(predict(mod_Zzi1, type = "prob")[, 1]))

## [1] 1.00019

cbind(Poisson = AIC(mod_Zpoiss),
      NegBin  = AIC(mod_Znb),
      Hurdle  = AIC(mod_Zhurd1),
      ZeroInf = AIC(mod_Zzi1))

##       Poisson   NegBin   Hurdle  ZeroInf
## [1,] 1495.434 1446.461 1415.743 1414.922

tab <- rbind(Poisson = c(mod_Zpoiss$coefficients, NA, NA),
             NegBin = c(mod_Znb$coefficients, NA, NA),
             Hurdle = unlist(mod_Zhurd1$coefficients),  ## NOTE: the hurdle part predicts positive counts, not zeros!!
             ZeroInfl = unlist(mod_Zzi1$coefficients),  ## NOTE: the binary part predicts zeros, not counts!!
             Truth = c(attr(AliensZ, "param.eta")$eggs, attr(AliensZ, "param.eta")$happy))
colnames(tab) <- c("Int.count", "Slope.count", "Int.bin", "Slope.bin")
tab

##           Int.count Slope.count   Int.bin  Slope.bin
## Poisson  -3.3117027   0.2520928        NA         NA
## NegBin   -3.4503702   0.2659359        NA         NA
## Hurdle   -1.0563084   0.1094210 -3.876696  0.3068416
## ZeroInfl -0.9865184   0.1033290  2.916296 -0.2869336
## Truth    -1.0000000   0.1000000 -3.000000  0.3000000

What is it?

Separation occurs when a level or combination of levels for categorical predictor, or when a particular threshold along a continuous predictor, predicts the outcomes perfectly.

Complete or quasi separation?

From Wikipedia:

For example, if the predictor X is continuous, and the outcome y = 1 for all observed x > 2. If the outcome values are perfectly determined by the predictor (e.g., y = 0 when x ≤ 2) then the condition "complete separation" is said to occur. If instead there is some overlap (e.g., y = 0 when x < 2, but y has observed values of 0 and 1 when x = 2) then "quasi-complete separation" occurs. A 2 × 2 table with an empty cell is an example of quasi-complete separation.

Consequences

• sometimes the model cannot be fitted
• when it does fit, the estimates and standard errors are usually wrong!

set.seed(1L)
n <- 50
test <- data.frame(happy = rbinom(2*n, prob = c(rep(0, n), rep(0.75, n)), size = 1), 
                   sp = factor(c(rep("sp1", n), rep("sp2", n))))
table(test$happy, test$sp)

##    
##     sp1 sp2
##   0  50  13
##   1   0  37

mod <- glm(happy ~ sp, data = test, family = binomial())

Let's tweak the data in a conservative way

test$happy[1] <- 1
table(test$happy, test$sp)

##    
##     sp1 sp2
##   0  49  13
##   1   1  37

mod2 <- glm(happy ~ sp, data = test, family = binomial())

summary(mod2)

## 
## Call:
## glm(formula = happy ~ sp, family = binomial(), data = test)
## 
## Deviance Residuals: 
##    Min      1Q  Median      3Q     Max  
## -1.641  -0.201  -0.201   0.776   2.797  
## 
## Coefficients:
##             Estimate Std. Error z value Pr(>|z|)    
## (Intercept)   -3.892      1.010  -3.853 0.000117 ***
## spsp2          4.938      1.060   4.657 3.21e-06 ***
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## (Dispersion parameter for binomial family taken to be 1)
## 
##     Null deviance: 132.81  on 99  degrees of freedom
## Residual deviance:  67.11  on 98  degrees of freedom
## AIC: 71.11
## 
## Number of Fisher Scoring iterations: 6

A solution for the binary case (not for binomial in general)

test$eggs[1] <- 0  ## restore the original data
library(safeBinaryRegression)  ## overload the glm function
mod3 <- glm(happy ~ sp, data = test, family = binomial())
summary(mod3)$coef

##              Estimate Std. Error   z value     Pr(>|z|)
## (Intercept) -3.891820   1.010091 -3.852941 1.167076e-04
## spsp2        4.937789   1.060299  4.656978 3.208842e-06

AIC(mod3)

## [1] 71.1096

A solution for the binary case (not for binomial in general)

library(spaMM)  ## with package e1071 installed!
mod4 <- fitme(happy ~ sp, data = test, family = binomial())
summary(mod4)

## formula: happy ~ sp
## Estimation of fixed effects by ML.
## Family: binomial ( link = logit ) 
##  ------------ Fixed effects (beta) ------------
##             Estimate Cond. SE t-value
## (Intercept)   -3.892     1.01  -3.853
## spsp2          4.938     1.06   4.657
##  ------------- Likelihood values  -------------
##                        logLik
## p(h)   (Likelihood): -33.5548

AIC(mod4)

A solution for the binary case (as well as for binomial in general!)

library(brglm)  ## there is a new version in development brglm2 but it still has issues
mod5 <- brglm(happy ~ sp, data = test, family = binomial())
summary(mod5)

## 
## Call:
## brglm(formula = happy ~ sp, family = binomial(), data = test)
## 
## 
## Coefficients:
##             Estimate Std. Error z value Pr(>|z|)    
## (Intercept)  -3.4965     0.8370  -4.177 2.95e-05 ***
## spsp2         4.5182     0.8963   5.041 4.63e-07 ***
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## (Dispersion parameter for binomial family taken to be 1)
## 
##     Null deviance: 124.46  on 99  degrees of freedom
## Residual deviance:  67.29  on 98  degrees of freedom
## Penalized deviance: 64.6585 
## AIC:  71.29

AIC(mod5)

## [1] 71.28971

• that many residuals can be computed for GLM and that they are often useless
• that computing residuals by parametric bootstraps in the way out
• that traditional goodness of fit tests are bad
• that most assumptions behing GLM are similar to those for LM
• that General Additive Models (GAM) can be useful to improve linearity
• how to diagnose and tackle overdispersion, zero-augmentation and separation

• 3.0 Introduction
• 3.1 Intervals & Tests
• 3.2 Residuals & Assumptions
• 3.3 Let's practice more